U.S. patent number 10,558,844 [Application Number 15/845,695] was granted by the patent office on 2020-02-11 for lightweight 3d vision camera with intelligent segmentation engine for machine vision and auto identification.
This patent grant is currently assigned to DATALOGIC IP TECH S.r.l.. The grantee listed for this patent is Datalogic IP Tech S.r.l.. Invention is credited to Davide Benghi, Alessandro Brunetta, Francesco D'Ercoli, Silvia Marinoni.
![](/patent/grant/10558844/US10558844-20200211-D00000.png)
![](/patent/grant/10558844/US10558844-20200211-D00001.png)
![](/patent/grant/10558844/US10558844-20200211-D00002.png)
![](/patent/grant/10558844/US10558844-20200211-D00003.png)
![](/patent/grant/10558844/US10558844-20200211-D00004.png)
United States Patent |
10,558,844 |
D'Ercoli , et al. |
February 11, 2020 |
Lightweight 3D vision camera with intelligent segmentation engine
for machine vision and auto identification
Abstract
Various embodiments of the invention are implemented for an
entry level, compact and lightweight single package apparatus
combining a conventional high-resolution two-dimensional (2D)
camera with a low-resolution three-dimensional (3D) depth image
camera, capable to learn, through depth information, how to improve
the performance of a set of 2D identification and machine vision
algorithms in terms of speed-up (e.g. through regions of interests
(ROIs)) and raw discriminative power. The cameras simultaneously
capture images that are processed by an Intelligent Segmentation
Engine in the system to facilitate object recognition.
Inventors: |
D'Ercoli; Francesco (Bologna,
IT), Brunetta; Alessandro (Padua, IT),
Marinoni; Silvia (Castenaso, IT), Benghi; Davide
(Monte San Pietro, IT) |
Applicant: |
Name |
City |
State |
Country |
Type |
Datalogic IP Tech S.r.l. |
Lippo di Calderara di Reno |
N/A |
IT |
|
|
Assignee: |
DATALOGIC IP TECH S.r.l. (Lippo
di Calderara di, IT)
|
Family
ID: |
64744492 |
Appl.
No.: |
15/845,695 |
Filed: |
December 18, 2017 |
Prior Publication Data
|
|
|
|
Document
Identifier |
Publication Date |
|
US 20190188451 A1 |
Jun 20, 2019 |
|
Current U.S.
Class: |
1/1 |
Current CPC
Class: |
G06T
7/75 (20170101); G06K 9/00214 (20130101); G06K
9/6211 (20130101); G06K 9/3233 (20130101); G06K
9/4638 (20130101); G06F 16/5838 (20190101); G06K
9/00208 (20130101); G06T 2207/20212 (20130101); G06K
2209/19 (20130101) |
Current International
Class: |
G06K
9/00 (20060101); G06T 7/73 (20170101); G06F
16/583 (20190101); G06K 9/46 (20060101); G06K
9/62 (20060101); G06K 9/32 (20060101) |
References Cited
[Referenced By]
U.S. Patent Documents
Foreign Patent Documents
|
|
|
|
|
|
|
103033145 |
|
Apr 2013 |
|
CN |
|
2016167947 |
|
Oct 2016 |
|
WO |
|
2017056035 |
|
Apr 2017 |
|
WO |
|
Other References
US 9,721,135 B2, 08/2017, Todeschini et al. (withdrawn) cited by
applicant .
Ghobadi, (Seyed Eghbal. "Real time object recognition and tracking
using 2D/3D images." Date Sep. 16, 2010) (Year: 2010). cited by
examiner .
Tan, O et al., "Review: Segmentation and classification methods of
3D medical images," PubMed Commons, Mar. 2002; 26(3): 197-206.
https://www.ncbi.nlm.nih.gov/pubmed/16104307. cited by applicant
.
Basic Thresholding Operations--OpenCV v2.4.2 documentation,
Accessed Dec. 9, 2017 at:
http://docs.opencv.org/2.4.2/doc/tutorials/imgproc/threshold/threshold.ht-
ml. cited by applicant .
Key Specifications epc660, 3D TOF Imager 320 x 240 Pixel, 2013
ESPROS Photonics Corporation, TS_epc660_Key_Specitications,
www.espros.ch. cited by applicant .
OpenCV: OpenCV modules, Accessed Dec. 9, 2017 at:
https://docs.opencv.org/3.1.0/ cited by applicant .
TOF 3D Cameras
(/EN/PRODUCTS/TOF-3D-CAMERAS/CATEGORY/TOF-3D-CAMERAS/) Accessed
Dec. 9, 2017 at:
https://www.bluetechnix.com/en/products/tof-3D-cameras. cited by
applicant .
Tangelder, Johan W.H. et al, "A survey of content based 3D shape
retrieval methods," Published online: Dec. 8, 2007, Multimed Tools
Appl (2008) 39:441-471, DOI 10.1007/s11042-007-0181-0
http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.463.7418&rep=rep-
1&type=pdf. cited by applicant .
Lang, Dagmar et al., "Semantic mapping for mobile outdoor robots,"
Machine Vision Applications (MVA), 2015 14th IAPR International
Conference, May 18-22, 2015. Available at:
http://ieeexplore.ieee.org/xpl/login.jsp?tp=&arnumber=7153196&url=http%3A-
%2F%2Fieeexplore.ieee.org%2Fxpls%2Fabs_all.jsp%3Farnumber%3D7153196.
cited by applicant .
TriSpector 1000 3D Vision Operating Instructions, Sep. 19, 2017.
Available at:
https://www.sick.com/media/dox/7/67/367/Operating_instructions_TriSpe-
ctor1000_3D_Vision_en_IM0060367.PDF. cited by applicant .
Extended search report dated Apr. 24, 2019 in European Patent
Application No. 18213164.9, 7 pages. cited by applicant.
|
Primary Examiner: Garcia; Santiago
Attorney, Agent or Firm: Shook, Hardy & Bacon,
L.L.P.
Claims
Having thus described the invention, what is claimed is:
1. A method of object recognition in a hybrid 2D/3D vision system,
comprising: acquiring a three-dimensional image of an object into a
vision device, wherein the vision device has a processor coupled to
a memory; acquiring one or more two-dimensional images of the
object into the vision device, wherein the three-dimensional image
is associated with the one or more two-dimensional images;
identifying the three-dimensional image through three-dimensional
shape recognition, wherein identifying the three-dimensional image
comprises comparing one or more shapes in the three-dimensional
image to objects stored in a database, and wherein a union of the
one or more shapes generates the three-dimensional image; if a
member of the one or more shapes of the three-dimensional image
matches to one of the objects in the database, associating the one
of the objects to a region of interest from the one or more
two-dimensional images; simultaneously, analyzing the member of the
one or more shapes of the three-dimensional image that matches to
the one of the objects and the region of interest from the one or
more two-dimensional images; and detecting a new region of interest
from the one or more two-dimensional images if another member of
the one or more shapes of the three-dimensional image does not
match to another one of the objects in the database; and increasing
a counter at the object responsive to a shape descriptor of the
object being compared and matched to the one of the other shape
descriptors in the database, wherein a highest value of the counter
of the object indicates the strongest reinforcement over the other
objects in the database with a lower value of the counter.
2. The method of claim 1, wherein the new region of interest is
stored in the database or another database.
3. The method of claim 1, wherein the database is a local
database.
4. A method of object recognition in a hybrid 2D/3D vision system,
comprising: acquiring a three-dimensional image of an object into a
vision device, wherein the vision device has a processor coupled to
a memory; acquiring one or more two-dimensional images of the
object into the vision device, wherein the three-dimensional image
is associated with the one or more two-dimensional images;
computing shape descriptors for the object; comparing the shape
descriptors for the object to other shape descriptors that are
stored in a database; if a shape descriptor of the object matches
one of the other shape descriptors in the database, designating the
object as an old object that is already in the database, wherein
the object matches the old object in the database; if the shape
descriptor of the object does not match one of the other shape
descriptors in the database, designating the object as a new object
and storing the object in the database; where the object matches
the old object in the database, assigning a weight factor or
reinforcement to the object; increasing the weight factor or
reinforcement of the object each time the shape descriptor of the
object is used in a comparison and matches to the one of the other
shape descriptors, wherein the object that is more frequently
classified has a stronger reinforcement; and for a comparison that
is not sufficient, using the object that has a strongest
reinforcement over other objects in the database with lesser
reinforcements.
5. The method of claim 4, further comprising updating the object
upon a subsequent comparison and the object is already in the
database designated as the old object or the new object.
6. The method of claim 4, wherein the shape descriptors are the
contours or shapes of items that are described.
7. A compact and portable device, comprising: a high-resolution
two-dimensional (2D) camera configured to capture 2D image; a
low-resolution three-dimensional (3D) camera configured to capture
a 3D image, wherein the 3D image is mapped into the 2D image to
form a calibrated image; and a segmentation engine configured to:
compute shape descriptors for the objects; compare the shape
descriptors for the object to other shape descriptors that are
stored in a database and in response thereto; designate the object
as a new object and storing the object in the database if the shape
descriptor of the object does not match one of the other shape
descriptors in the database; assign reinforcement information to
the shape descriptors stored in the database; and adjust the
reinforcement information of at least one of the shape descriptors
stored in the database each time the respective shape descriptor is
determined to match the shape descriptor of the object being
analyzed, wherein the reinforcement information is a weight factor,
and wherein the segmentation engine is configured to compare the
shape descriptors for the object to other shape descriptors that
are stored in a database by first searching for shape descriptors
stored in the database that have higher weight factors before
searching for shape descriptors having lower weight factors.
8. The device of claim 7, wherein the low-resolution 3D camera has
reduced 3D computational overload and power consumption.
9. The device of claim 7, wherein the high-resolution 2D camera and
the low-resolution 3D camera have sensors optimized for the
infrared (IR) range.
10. The device of claim 7, wherein the high-resolution 2D camera
has a sensor optimized for the visible range and the low-resolution
3D camera have a sensor optimized for the infrared (IR) range.
11. The device of claim 7, wherein the high-resolution 2D camera
and the low-resolution 3D camera simultaneously frame-grab an
image.
12. The method of claim 6, further comprising applying a threshold
to the 3D image to discard background objects in the 3D image.
13. The device of claim 7, wherein the 3D camera includes a
time-of-light sensor.
14. The device of claim 7, further comprising a lighting system
shared by the 2D camera and 3D camera.
15. The device of claim 7, wherein the 2D camera and 3D camera are
mounted together to have parallel optical axes, wherein a first
sensor of the 2D camera and a second sensor of the 3D camera point
in a same direction and have similar viewing areas.
16. The device of claim 15, wherein the 3D camera is mounted on top
of the 2D camera within a common geometrical environment.
17. The device of claim 7, wherein the weight factor is based on a
counter being increased each time the respective shape descriptor
is determined to match the shape descriptor of the object being
analyzed.
18. The method of claim 1, wherein the region of interest is
defined by borders of the object from the two-dimensional image.
Description
CROSS-REFERENCE TO RELATED APPLICATIONS
None.
BACKGROUND OF THE INVENTION
Three-dimensional (3D) imaging is a fast growing trend, especially
in the Industrial Automation field, where embedded vision sensors
are widely used. Range imaging techniques are spreading not only
due to Stereoscopic or Structured Light approaches, but also with
new image sensors designed to measure the pixel distance by various
methods and technologies, such as time-of-flight (TOF) and
light-field (aka plenoptic) cameras.
A relevant example of an industrial, entry-level apparatus (based
on different technology and method) is the TriSpector 1000 3D
Vision Family manufactured by SICK AG of Waldkirch, Germany.
Although a single optimal approach has not yet emerged, the
possibility to equip a 3D camera with soft Artificial Intelligence
is a very hot topic in the pattern recognition and machine vision
community.
BRIEF SUMMARY OF THE INVENTION
An entry level, compact and lightweight single package apparatus
combining a conventional two-dimensional (2D) camera with a
three-dimensional (3D) depth image camera, capable to learn,
through depth information, how to improve the performance of a set
of 2D identification and machine vision algorithms in terms of
speed-up (e.g. through regions of interests (ROIs)) and raw
discriminative power. This is achieved by implementing an
"Intelligent Segmentation Engine" (ISE) consisting of an expert
system capable of learning and building an incremental history of
the known objects, while receiving appropriate feedback from 2D
algorithms.
In a first aspect, a method of object recognition in a hybrid 2D/3D
vision system is provided that includes acquiring a
three-dimensional image of an object into a vision device. The
vision device has a processor coupled to a memory. Two-dimensional
images of the object are acquired into the vision device. The
three-dimensional image is associated with the two-dimensional
images. The three-dimensional image is identified through
three-dimensional shape recognition. The three-dimensional image is
identified by comparing shapes in the three-dimensional image to
objects in a database. A union of the shapes generates the
three-dimensional image. If a member of the shapes of the
three-dimensional image matches to one of the objects in the
database, that one of the objects is matched to a region of
interest from the two-dimensional images. Simultaneously, the
member of the shapes of the three-dimensional image that matches to
the one of the objects is matched to the region of interest from
the two-dimensional images to speed up an object recognition
process. A new region of interest from the two-dimensional images
is detected if another member of the shapes of the
three-dimensional image does not match to another one of the
objects in the database. The new region of interest matches to
another member of the shapes of the three-dimensional image.
In another aspect, a method of object recognition in a hybrid 2D/3D
vision system is provided that includes acquiring a
three-dimensional image of an object into a vision device. The
vision device has a processor coupled to a memory. Two-dimensional
images of the object are acquired into the vision device. The
three-dimensional image is associated with the two-dimensional
images. Shape descriptors are computed for the object. The shape
descriptors for the object are compared to other shape descriptors
that are stored in a database. If a shape descriptor of the object
matches one of the other shape descriptors in the database, the
object is designated as an old object that is already in the
database. The object matches the old object in the database. If the
shape descriptor of the object does not match one of the other
shape descriptors in the database, the object is designated as a
new object and the object is stored in the database. Where the
object matches the old object in the database, a weight factor or
reinforcement is assigned to the object. The weight factor or
reinforcement of the object is increased each time the shape
descriptor of the object is used in a comparison and matches to one
of the other shape descriptors. The object that is more frequently
classified has a stronger reinforcement. For a comparison that is
not sufficient, the object that has a strongest reinforcement is
used over other objects in the database with lesser
reinforcements.
In yet another aspect, a compact and portable device is provided
that includes several components. A high-resolution two-dimensional
(2D) camera is included in the device. A low-resolution
three-dimensional (3D) camera is included in the device. A lighting
system is shared by the 2D camera and 3D camera. The 2D camera and
3D camera are mounted together to have parallel optical axes. A
first sensor of the 2D camera and a second sensor of the 3D camera
point in a same direction and have similar viewing areas. The 3D
camera is mounted on top of the 2D camera. The first sensor and the
second sensor simultaneously capture a 2D image and 3D image of a
same viewing area. The 3D image is mapped into the 2D image to form
a composite image.
BRIEF DESCRIPTION OF THE SEVERAL VIEWS OF THE DRAWING
FIG. 1 illustrates an exemplary system combining 2D and 3D cameras
for machine vision and identification, in an implementation of an
embodiment of the invention;
FIG. 2 illustrates an exemplary device combining 2D and 3D cameras,
in an implementation of an embodiment of the invention;
FIG. 3 illustrates an exemplary process of an Intelligent
Segmentation Engine (ISE), in an implementation of an embodiment of
the invention; and
FIG. 4 illustrates an exemplary operational overview of combined 2D
and 3D cameras for machine vision and identification, in an
implementation of an embodiment of the invention.
DETAILED DESCRIPTION OF THE INVENTION
The proposed apparatus and method can be used in automatic
identification and machine vision applications, where there is the
need for a low-cost solution without depth accuracy requirements
mandatory in case of volume dimensioning or body scanning.
Combining a low cost 3D-TOF approach with a 2D high-resolution
camera in a single package sensor together with an original machine
learning segmentation method has some distinctive advantages:
Additional performance and speed-up for inspection and
identification (barcode, optical character recognition (OCR)).
Reduced total cost, size and weight for a 3D industrial camera
Reduced computational load involved in 3D image processing (by mean
of objects (OBJs) localization) Reduced memory needs to store 3D
images (3D compression: relevant OBJs only) Reduced network
bandwidth needs when a 3D image has to be transmitted
Hybrid 2D/3D intelligent vision systems will be interesting in the
next few years to achieve high performances not only in inspection
and automatic identification but also in new applications such as:
Gesture recognition for self-shop applications, where high 3D
accuracy is not required and the cost is a key-parameter Robotic
barcode scanning Augmented Reality rendering: a framework for
interactive AR applications, where the depth map is necessary for
visual or physical interaction between synthetic and real,
identified objects.
Apparatus
An embodiment of the invention consists of a compact, single
package vision sensor 200 containing a standard high-resolution 2D
camera 210 (brightness image) and a lower resolution 3D camera 220
(depth image) pair, mounted and calibrated in an optical and
geometrical environment 230. The resulting, combined camera system
200, is a compact and portable device, with reduced 3D
computational overload and power consumption. The combined camera
system 200 includes a lighting system.
In a preferred embodiment, system 200 will be shared by the two
sensors 210 and 220 (both the sensors optimized for infrared (IR)
range), but system 200 could also be composed of two different
sensor types (IR-TOF sensor combined with 2D visible range sensor).
Even if TOF cameras have less depth accuracy compared to other
types of 3D sensors (e.g. laser scanners, structured light, or
stereo vision systems), this is not a limitation for the intended
purposes of the proposed invention, since it can benefit from the
aid of a high-resolution 2D sensor. On the contrary, in respect to
other 3D technologies, TOF cameras can be fitted to real-time
applications where automatic acquisition of 3D models involves
sensor movement.
To maximize the performance of the two-camera combination in system
200, the cameras 210 and 220 should be placed as close as possible
to each other. It is proposed that the 3D TOF camera 220 be mounted
on top of the 2D camera 210. Moreover, the two sensors 210 and 220
should have appropriate optical systems in order to share similar
viewing areas. In this way, the 3D information can be matched to or
mapped inside the 2D image. From an operational point of view, the
image is frame grabbed simultaneously by both the cameras 210 and
220 and then the TOF image from camera 220 is mapped into the 2D
image of camera 210 (a threshold can be applied to the depth image
to discard background objects). A TOF sensor can be an industrial,
entry level low-cost IR-based depth camera, with a low resolution
(320.times.240 pixels or QVGA).
The TOF camera 220 and the 2D camera 210 are calibrated, and a
mapping is made from the depth data to the 2D image. Calibration of
2D and 3D images is a state-of-the-art process, quite similar to a
stereo camera calibration problem. It is only a matter of starting
from 3D coordinates of the TOF camera 220 to find the corresponding
2D coordinates instead of building 3D coordinates from two sets of
2D coordinates. Given the calibrated image, segmentation will be
performed directly on the depth image from the TOF camera 220,
without using 2D information from the 2D camera 210. This is a
fundamental task for 3D vision, often the very first step. It
consists of partitioning the corresponding digital image into
groups of pixels (regions) with similar characteristics (e.g.
color, intensity, texture, etc.) in order to simplify the detection
and classification of objects in the scene.
Through segmentation, different objects in the scene are delimited
by contours or segments consisting of boundaries such as lines and
curves (closely related to edges). In addition, 3D image
segmentation allows for subsequent 3D rendering of multiple objects
and a quantitative analysis for their size, density and other
parameters. Edges identified by edge detection are often
disconnected while object segmentation needs closed region
boundaries. There is no overlap between different regions and the
union of all the segmented regions generates the entire image.
Segmentation can also be applied to a stack of images (this is
typical in medical imaging) so that the resulting contours can
allow for 3D reconstructions through a kind of interpolation
algorithm (e.g. Marching cubes). Many algorithms and techniques
have been proposed for image segmentation, especially in the
biomedical field. Particularly, the selected technique must be
matched with the application's specific knowledge to get the best
results.
Intelligent Segmentation Engine
An important aspect of the invention is the custom Intelligent
Segmentation Engine (ISE) 300 provided for the system 200. It
consists of five main components, depth threshold computing 310,
labeling 320, classification 330, assistant module 340, and objects
350, designed to achieve automatic segmentation and classification
while providing "expert information" to 2D algorithms.
(1) Depth Threshold Computing
A set of depth thresholds are computed (310) in order to extract
labeled regions (e.g. foreground, background, etc.) from the depth
image of camera 220. It should be noted that a unique global
threshold generally fails to detect all objects. Many known
algorithms are available for this task.
(2) Labeling
For all the computed depth thresholds, some objects are extracted
and labeled in 320.
(3) Classification
For all different extracted objects, the following processing steps
are performed as part of classification 330. Robust 3D geometrical
shape descriptors are computed in order to get a meaningful
classification metric. Shape descriptors of the current object are
compared to those in already available records stored in a local or
remote data base. Based on this comparison, the current object can
be said "new OBJ" (never seen before) or "old OBJ" (already seen
before). An additional information called "reinforcement" is
associated to every object in the data base. It is used as a weight
factor in the previous comparison: objects with stronger
reinforcement (i.e. more frequently classified) will be preferred
in case of doubtful comparisons. It could also be possible to
define a set of rules in order to better exploit this information
to improve future object classifications.
(4) Assistant Module (AM)
Classified objects can be either new objects or object already
found in past images as determined classification 330. For both new
and old OBJs, a set of specific 2D ROIs, corresponding to currently
recognized 3D objects, are set as actual input for 2D algorithms.
In case of "old OBJ", together with the ROIs, some additional
"expert information" about the expected output identification or
inspection can be supplied to the 2D algorithms to facilitate their
operations. A ROI is typically defined as the borders of an object
or the polygonal shape of an item in a 2D image. An example of the
types of information discussed above is provided in TAB 1.
TABLE-US-00001 TABLE 1 EXPERT_INFORMATION ROIS Regions of interest
defined as one or more irregular 2D image sub-regions corresponding
to one or more classified OBJs EXP_POS For every ROI, the expected
(probable) location of all (old OBJ only) items already identified
in previous object occurrences (past acquisitions). EXP_SIZE For
every ROI, the expected (probable) size of all (old OBJ only) items
already identified in previous object occurrences (past
acquisitions). EXP_MODULE For every ROI, the expected (probable)
barcode (old OBJ only) module (mm/mil) of all items already
identified in previous object occurrences (past acquisitions).
Once the AM 340 receives feedback from 2D algorithms (e.g.
identified codes, inspection results), it will be able to: Enrich
the information associated to the objects with the current items
identified (new or old objects), by increasing some counters
corresponding to the "probabilities" referred in TAB 1, to be
stored with the object. Store the objects in the database: from
scratch if "new," updated with reinforcement if "old."
Reinforcement consists of a counter to be increased after each new
classification of the object.
(5) Objects Database
The object database 350 can be viewed as a declarative memory
referred to as general "world knowledge" accumulated through both
3D segmentation, 2D identification, and inspection experience. It
is important to know "what an object is," not only to know that "an
object exists."
Every OBJ record consists of some fields, as reported in TAB 2,
which precisely defines what the object is, what (items) is on its
surface, and how much is common (hence, easy to find).
As an option: Additional information called "relations" can be
associated to every object in the database (350) to detail its
relation with other objects (inclusion, proximity, similarity,
etc.). It could be used to facilitate the classification task.
TABLE-US-00002 TABLE 2 OBJECT_RECORD 3D SHAPE May be a polygon mesh
as a collection of vertices, DESCRIPTORS edges and faces able to
define the shape of a polyhedral object in 3D solid modeling (faces
usually consist of triangles). ITEMS Identified patterns (barcodes,
OCR text strings, inspection symbols) on the object, with related
"expert information" (e.g. positions, sizes and modules).
ENRICHMENT Number of times the items has been identified
REINFORCEMENT Number of times the object has been classified
In FIG. 4, a concept overview 400 is shown. An image 410 is
captured with cameras 210 and 220 and processed with ISE 300.
A very important point of the invention lies in its ability to
enable real-time processing of actual complex scenes. Particularly,
a brute force approach that uses high-resolution 3D images (e.g. by
stereo cameras or structured light) associated with high
computational-load machine learning methods (e.g. deep
convolutional neural networks, Bayesian networks, etc.) would
probably require more processing time and specific hardware
resources (such as dedicated GPUs) that are not compatible with the
invention, in low cost automatic identification and machine vision
applications.
More specifically, embodiments of the invention use the following
elements in a compact package:
A Low Resolution 3D TOF Camera
Since the depth image involves a small number of data points (e.g.
320.times.240 pixels, QVGA), computing depth thresholds to extract
labeled regions and computing raw shape descriptors (there is no
need to model any object detail) can be considered a low
computational-load task, feasible in the embedded environment.
A High Resolution 2D Camera
The maximum benefits in terms of processing time can be achieved
when using a high-resolution 2D imager (e.g. sensors with more than
5 megapixels). This is because a low resolution 3D camera
effectively filter information to 2D algorithms.
A Simple Reinforcement/Enrichment Approach
Instead of a huge machine learning method, collected
EXPERT_INFORMATION is used to be linked with OBJECT_RECORDs that
can be managed by means of simple row/column tables. In other
words, the most "reinforced" objects will be the first to be
searched during the classification phase (and it will be first
searched in the most likely positions).
Limited/Prioritized 2D Image to be Processed
A further processing time saving is due to the fact that 2D
identification and machine vision algorithms can take advantage of
EXPERT_INFORMATION to analyze first (or even exclusively) the
region of image associated with the segmented 3D object. In
addition, the algorithm parameters can be automatically adjusted
over that region in agreement with the expected items that will
most likely be found (this information is stored as
"enrichment").
Many different arrangements of the various components depicted, as
well as components not shown, are possible without departing from
the scope of the claims below. Embodiments of our technology have
been described with the intent to be illustrative rather than
restrictive. Alternative embodiments will become apparent to
readers of this disclosure after and because of reading it.
Alternative means of implementing the aforementioned invention can
be completed without departing from the scope of the claims below.
Certain features and sub-combinations are of utility and may be
employed without reference to other features and sub-combinations
and are contemplated within the scope of the claims.
* * * * *
References